2 * SIPR / ACELP.NET decoder
4 * Copyright (c) 2008 Vladimir Voroshilov
5 * Copyright (c) 2009 Vitor Sessak
7 * This file is part of FFmpeg.
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
27 #include "libavutil/mathematics.h"
29 #define ALT_BITSTREAM_READER_LE
34 #include "celp_math.h"
35 #include "acelp_vectors.h"
36 #include "acelp_pitch_delay.h"
37 #include "acelp_filters.h"
38 #include "celp_filters.h"
40 #define MAX_SUBFRAME_COUNT 5
46 const char *mode_name;
47 uint16_t bits_per_frame;
48 uint8_t subframe_count;
49 uint8_t frames_per_packet;
50 float pitch_sharp_factor;
52 /* bitstream parameters */
53 uint8_t number_of_fc_indexes;
54 uint8_t ma_predictor_bits; ///< size in bits of the switched MA predictor
56 /** size in bits of the i-th stage vector of quantizer */
57 uint8_t vq_indexes_bits[5];
59 /** size in bits of the adaptive-codebook index for every subframe */
60 uint8_t pitch_delay_bits[5];
62 uint8_t gp_index_bits;
63 uint8_t fc_index_bits[10]; ///< size in bits of the fixed codebook indexes
64 uint8_t gc_index_bits; ///< size in bits of the gain codebook indexes
67 static const SiprModeParam modes[MODE_COUNT] = {
70 .bits_per_frame = 160,
71 .subframe_count = SUBFRAME_COUNT_16k,
72 .frames_per_packet = 1,
73 .pitch_sharp_factor = 0.00,
75 .number_of_fc_indexes = 10,
76 .ma_predictor_bits = 1,
77 .vq_indexes_bits = {7, 8, 7, 7, 7},
78 .pitch_delay_bits = {9, 6},
80 .fc_index_bits = {4, 5, 4, 5, 4, 5, 4, 5, 4, 5},
86 .bits_per_frame = 152,
88 .frames_per_packet = 1,
89 .pitch_sharp_factor = 0.8,
91 .number_of_fc_indexes = 3,
92 .ma_predictor_bits = 0,
93 .vq_indexes_bits = {6, 7, 7, 7, 5},
94 .pitch_delay_bits = {8, 5, 5},
96 .fc_index_bits = {9, 9, 9},
102 .bits_per_frame = 232,
104 .frames_per_packet = 2,
105 .pitch_sharp_factor = 0.8,
107 .number_of_fc_indexes = 3,
108 .ma_predictor_bits = 0,
109 .vq_indexes_bits = {6, 7, 7, 7, 5},
110 .pitch_delay_bits = {8, 5, 5},
112 .fc_index_bits = {5, 5, 5},
118 .bits_per_frame = 296,
120 .frames_per_packet = 2,
121 .pitch_sharp_factor = 0.85,
123 .number_of_fc_indexes = 1,
124 .ma_predictor_bits = 0,
125 .vq_indexes_bits = {6, 7, 7, 7, 5},
126 .pitch_delay_bits = {8, 5, 8, 5, 5},
128 .fc_index_bits = {10},
133 const float ff_pow_0_5[] = {
134 1.0/(1 << 1), 1.0/(1 << 2), 1.0/(1 << 3), 1.0/(1 << 4),
135 1.0/(1 << 5), 1.0/(1 << 6), 1.0/(1 << 7), 1.0/(1 << 8),
136 1.0/(1 << 9), 1.0/(1 << 10), 1.0/(1 << 11), 1.0/(1 << 12),
137 1.0/(1 << 13), 1.0/(1 << 14), 1.0/(1 << 15), 1.0/(1 << 16)
140 static void dequant(float *out, const int *idx, const float *cbs[])
146 for (i = 0; i < num_vec; i++)
147 memcpy(out + stride*i, cbs[i] + stride*idx[i], stride*sizeof(float));
151 static void lsf_decode_fp(float *lsfnew, float *lsf_history,
152 const SiprParameters *parm)
155 float lsf_tmp[LP_FILTER_ORDER];
157 dequant(lsf_tmp, parm->vq_indexes, lsf_codebooks);
159 for (i = 0; i < LP_FILTER_ORDER; i++)
160 lsfnew[i] = lsf_history[i] * 0.33 + lsf_tmp[i] + mean_lsf[i];
162 ff_sort_nearly_sorted_floats(lsfnew, LP_FILTER_ORDER - 1);
164 /* Note that a minimum distance is not enforced between the last value and
165 the previous one, contrary to what is done in ff_acelp_reorder_lsf() */
166 ff_set_min_dist_lsf(lsfnew, LSFQ_DIFF_MIN, LP_FILTER_ORDER - 1);
167 lsfnew[9] = FFMIN(lsfnew[LP_FILTER_ORDER - 1], 1.3 * M_PI);
169 memcpy(lsf_history, lsf_tmp, LP_FILTER_ORDER * sizeof(*lsf_history));
171 for (i = 0; i < LP_FILTER_ORDER - 1; i++)
172 lsfnew[i] = cos(lsfnew[i]);
173 lsfnew[LP_FILTER_ORDER - 1] *= 6.153848 / M_PI;
176 /** Apply pitch lag to the fixed vector (AMR section 6.1.2). */
177 static void pitch_sharpening(int pitch_lag_int, float beta,
182 for (i = pitch_lag_int; i < SUBFR_SIZE; i++)
183 fixed_vector[i] += beta * fixed_vector[i - pitch_lag_int];
187 * Extract decoding parameters from the input bitstream.
188 * @param parms parameters structure
189 * @param pgb pointer to initialized GetBitContext structure
191 static void decode_parameters(SiprParameters* parms, GetBitContext *pgb,
192 const SiprModeParam *p)
196 parms->ma_pred_switch = get_bits(pgb, p->ma_predictor_bits);
198 for (i = 0; i < 5; i++)
199 parms->vq_indexes[i] = get_bits(pgb, p->vq_indexes_bits[i]);
201 for (i = 0; i < p->subframe_count; i++) {
202 parms->pitch_delay[i] = get_bits(pgb, p->pitch_delay_bits[i]);
203 parms->gp_index[i] = get_bits(pgb, p->gp_index_bits);
205 for (j = 0; j < p->number_of_fc_indexes; j++)
206 parms->fc_indexes[i][j] = get_bits(pgb, p->fc_index_bits[j]);
208 parms->gc_index[i] = get_bits(pgb, p->gc_index_bits);
212 static void sipr_decode_lp(float *lsfnew, const float *lsfold, float *Az,
215 double lsfint[LP_FILTER_ORDER];
217 float t, t0 = 1.0 / num_subfr;
220 for (i = 0; i < num_subfr; i++) {
221 for (j = 0; j < LP_FILTER_ORDER; j++)
222 lsfint[j] = lsfold[j] * (1 - t) + t * lsfnew[j];
224 ff_amrwb_lsp2lpc(lsfint, Az, LP_FILTER_ORDER);
225 Az += LP_FILTER_ORDER;
231 * Evaluate the adaptive impulse response.
233 static void eval_ir(const float *Az, int pitch_lag, float *freq,
234 float pitch_sharp_factor)
236 float tmp1[SUBFR_SIZE+1], tmp2[LP_FILTER_ORDER+1];
240 for (i = 0; i < LP_FILTER_ORDER; i++) {
241 tmp1[i+1] = Az[i] * ff_pow_0_55[i];
242 tmp2[i ] = Az[i] * ff_pow_0_7 [i];
244 memset(tmp1 + 11, 0, 37 * sizeof(float));
246 ff_celp_lp_synthesis_filterf(freq, tmp2, tmp1, SUBFR_SIZE,
249 pitch_sharpening(pitch_lag, pitch_sharp_factor, freq);
253 * Evaluate the convolution of a vector with a sparse vector.
255 static void convolute_with_sparse(float *out, const AMRFixed *pulses,
256 const float *shape, int length)
260 memset(out, 0, length*sizeof(float));
261 for (i = 0; i < pulses->n; i++)
262 for (j = pulses->x[i]; j < length; j++)
263 out[j] += pulses->y[i] * shape[j - pulses->x[i]];
267 * Apply postfilter, very similar to AMR one.
269 static void postfilter_5k0(SiprContext *ctx, const float *lpc, float *samples)
271 float buf[SUBFR_SIZE + LP_FILTER_ORDER];
272 float *pole_out = buf + LP_FILTER_ORDER;
273 float lpc_n[LP_FILTER_ORDER];
274 float lpc_d[LP_FILTER_ORDER];
277 for (i = 0; i < LP_FILTER_ORDER; i++) {
278 lpc_d[i] = lpc[i] * ff_pow_0_75[i];
279 lpc_n[i] = lpc[i] * ff_pow_0_5 [i];
282 memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem,
283 LP_FILTER_ORDER*sizeof(float));
285 ff_celp_lp_synthesis_filterf(pole_out, lpc_d, samples, SUBFR_SIZE,
288 memcpy(ctx->postfilter_mem, pole_out + SUBFR_SIZE - LP_FILTER_ORDER,
289 LP_FILTER_ORDER*sizeof(float));
291 ff_tilt_compensation(&ctx->tilt_mem, 0.4, pole_out, SUBFR_SIZE);
293 memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem5k0,
294 LP_FILTER_ORDER*sizeof(*pole_out));
296 memcpy(ctx->postfilter_mem5k0, pole_out + SUBFR_SIZE - LP_FILTER_ORDER,
297 LP_FILTER_ORDER*sizeof(*pole_out));
299 ff_celp_lp_zero_synthesis_filterf(samples, lpc_n, pole_out, SUBFR_SIZE,
304 static void decode_fixed_sparse(AMRFixed *fixed_sparse, const int16_t *pulses,
305 SiprMode mode, int low_gain)
311 for (i = 0; i < 3; i++) {
312 fixed_sparse->x[i] = 3 * (pulses[i] & 0xf) + i;
313 fixed_sparse->y[i] = pulses[i] & 0x10 ? -1 : 1;
318 for (i = 0; i < 3; i++) {
319 fixed_sparse->x[2*i ] = 3 * ((pulses[i] >> 4) & 0xf) + i;
320 fixed_sparse->x[2*i + 1] = 3 * ( pulses[i] & 0xf) + i;
322 fixed_sparse->y[2*i ] = (pulses[i] & 0x100) ? -1.0: 1.0;
324 fixed_sparse->y[2*i + 1] =
325 (fixed_sparse->x[2*i + 1] < fixed_sparse->x[2*i]) ?
326 -fixed_sparse->y[2*i ] : fixed_sparse->y[2*i];
334 int offset = (pulses[0] & 0x200) ? 2 : 0;
337 for (i = 0; i < 3; i++) {
338 int index = (val & 0x7) * 6 + 4 - i*2;
340 fixed_sparse->y[i] = (offset + index) & 0x3 ? -1 : 1;
341 fixed_sparse->x[i] = index;
347 int pulse_subset = (pulses[0] >> 8) & 1;
349 fixed_sparse->x[0] = ((pulses[0] >> 4) & 15) * 3 + pulse_subset;
350 fixed_sparse->x[1] = ( pulses[0] & 15) * 3 + pulse_subset + 1;
352 fixed_sparse->y[0] = pulses[0] & 0x200 ? -1 : 1;
353 fixed_sparse->y[1] = -fixed_sparse->y[0];
360 static void decode_frame(SiprContext *ctx, SiprParameters *params,
364 int subframe_count = modes[ctx->mode].subframe_count;
365 int frame_size = subframe_count * SUBFR_SIZE;
366 float Az[LP_FILTER_ORDER * MAX_SUBFRAME_COUNT];
368 float ir_buf[SUBFR_SIZE + LP_FILTER_ORDER];
369 float lsf_new[LP_FILTER_ORDER];
370 float *impulse_response = ir_buf + LP_FILTER_ORDER;
371 float *synth = ctx->synth_buf + 16; // 16 instead of LP_FILTER_ORDER for
376 memset(ir_buf, 0, LP_FILTER_ORDER * sizeof(float));
377 lsf_decode_fp(lsf_new, ctx->lsf_history, params);
379 sipr_decode_lp(lsf_new, ctx->lsp_history, Az, subframe_count);
381 memcpy(ctx->lsp_history, lsf_new, LP_FILTER_ORDER * sizeof(float));
383 excitation = ctx->excitation + PITCH_DELAY_MAX + L_INTERPOL;
385 for (i = 0; i < subframe_count; i++) {
386 float *pAz = Az + i*LP_FILTER_ORDER;
387 float fixed_vector[SUBFR_SIZE];
389 float pitch_gain, gain_code, avg_energy;
391 ff_decode_pitch_lag(&T0, &T0_frac, params->pitch_delay[i], t0_first, i,
392 ctx->mode == MODE_5k0, 6);
394 if (i == 0 || (i == 2 && ctx->mode == MODE_5k0))
397 ff_acelp_interpolatef(excitation, excitation - T0 + (T0_frac <= 0),
399 2 * ((2 + T0_frac)%3 + 1), LP_FILTER_ORDER,
402 decode_fixed_sparse(&fixed_cb, params->fc_indexes[i], ctx->mode,
403 ctx->past_pitch_gain < 0.8);
405 eval_ir(pAz, T0, impulse_response, modes[ctx->mode].pitch_sharp_factor);
407 convolute_with_sparse(fixed_vector, &fixed_cb, impulse_response,
411 (0.01 + ff_dot_productf(fixed_vector, fixed_vector, SUBFR_SIZE))/
414 ctx->past_pitch_gain = pitch_gain = gain_cb[params->gc_index[i]][0];
416 gain_code = ff_amr_set_fixed_gain(gain_cb[params->gc_index[i]][1],
417 avg_energy, ctx->energy_history,
418 34 - 15.0/(0.05*M_LN10/M_LN2),
421 ff_weighted_vector_sumf(excitation, excitation, fixed_vector,
422 pitch_gain, gain_code, SUBFR_SIZE);
424 pitch_gain *= 0.5 * pitch_gain;
425 pitch_gain = FFMIN(pitch_gain, 0.4);
427 ctx->gain_mem = 0.7 * ctx->gain_mem + 0.3 * pitch_gain;
428 ctx->gain_mem = FFMIN(ctx->gain_mem, pitch_gain);
429 gain_code *= ctx->gain_mem;
431 for (j = 0; j < SUBFR_SIZE; j++)
432 fixed_vector[j] = excitation[j] - gain_code * fixed_vector[j];
434 if (ctx->mode == MODE_5k0) {
435 postfilter_5k0(ctx, pAz, fixed_vector);
437 ff_celp_lp_synthesis_filterf(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
438 pAz, excitation, SUBFR_SIZE,
442 ff_celp_lp_synthesis_filterf(synth + i*SUBFR_SIZE, pAz, fixed_vector,
443 SUBFR_SIZE, LP_FILTER_ORDER);
445 excitation += SUBFR_SIZE;
448 memcpy(synth - LP_FILTER_ORDER, synth + frame_size - LP_FILTER_ORDER,
449 LP_FILTER_ORDER * sizeof(float));
451 if (ctx->mode == MODE_5k0) {
452 for (i = 0; i < subframe_count; i++) {
453 float energy = ff_dot_productf(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
454 ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
456 ff_adaptive_gain_control(&synth[i * SUBFR_SIZE],
457 &synth[i * SUBFR_SIZE], energy,
458 SUBFR_SIZE, 0.9, &ctx->postfilter_agc);
461 memcpy(ctx->postfilter_syn5k0, ctx->postfilter_syn5k0 + frame_size,
462 LP_FILTER_ORDER*sizeof(float));
464 memcpy(ctx->excitation, excitation - PITCH_DELAY_MAX - L_INTERPOL,
465 (PITCH_DELAY_MAX + L_INTERPOL) * sizeof(float));
467 ff_acelp_apply_order_2_transfer_function(out_data, synth,
468 (const float[2]) {-1.99997 , 1.000000000},
469 (const float[2]) {-1.93307352, 0.935891986},
471 ctx->highpass_filt_mem,
475 static av_cold int sipr_decoder_init(AVCodecContext * avctx)
477 SiprContext *ctx = avctx->priv_data;
480 if (avctx->bit_rate > 12200) ctx->mode = MODE_16k;
481 else if (avctx->bit_rate > 7500 ) ctx->mode = MODE_8k5;
482 else if (avctx->bit_rate > 5750 ) ctx->mode = MODE_6k5;
483 else ctx->mode = MODE_5k0;
485 av_log(avctx, AV_LOG_DEBUG, "Mode: %s\n", modes[ctx->mode].mode_name);
487 if (ctx->mode == MODE_16k)
488 ff_sipr_init_16k(ctx);
490 for (i = 0; i < LP_FILTER_ORDER; i++)
491 ctx->lsp_history[i] = cos((i+1) * M_PI / (LP_FILTER_ORDER + 1));
493 for (i = 0; i < 4; i++)
494 ctx->energy_history[i] = -14;
496 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
498 dsputil_init(&ctx->dsp, avctx);
503 static int sipr_decode_frame(AVCodecContext *avctx, void *datap,
504 int *data_size, AVPacket *avpkt)
506 SiprContext *ctx = avctx->priv_data;
507 const uint8_t *buf=avpkt->data;
509 const SiprModeParam *mode_par = &modes[ctx->mode];
512 int subframe_size = ctx->mode == MODE_16k ? L_SUBFR_16k : SUBFR_SIZE;
516 if (avpkt->size < (mode_par->bits_per_frame >> 3)) {
517 av_log(avctx, AV_LOG_ERROR,
518 "Error processing packet: packet size (%d) too small\n",
524 if (*data_size < subframe_size * mode_par->subframe_count * sizeof(float)) {
525 av_log(avctx, AV_LOG_ERROR,
526 "Error processing packet: output buffer (%d) too small\n",
533 init_get_bits(&gb, buf, mode_par->bits_per_frame);
535 for (i = 0; i < mode_par->frames_per_packet; i++) {
536 decode_parameters(&parm, &gb, mode_par);
538 if (ctx->mode == MODE_16k)
539 ff_sipr_decode_frame_16k(ctx, &parm, data);
541 decode_frame(ctx, &parm, data);
543 data += subframe_size * mode_par->subframe_count;
546 *data_size = mode_par->frames_per_packet * subframe_size *
547 mode_par->subframe_count * sizeof(float);
549 return mode_par->bits_per_frame >> 3;
552 AVCodec sipr_decoder = {
561 .long_name = NULL_IF_CONFIG_SMALL("RealAudio SIPR / ACELP.NET"),